Identification of Autism Spectrum Disorder Using Multiple Functional Connectivity-Based Graph Convolutional Network

Neurology Computer Science Information Science Artificial Intelligence Medicine Clinical Psychology Medical Imaging
autism spectrum disorder graph convolutional network functional connectivity brain network readout layer

The title of this paper is “Identification of Autism Spectrum Disorder Using Multiple Functional Connectivity-based Graph Convolutional Network,” published in the journal “Medical & Biological Engineering & Computing,” volume 62, pages 2133-2144, in 2024. This paper proposes a multiple functional connectivity-based graph convolutional network (mfc-GCN) framework by combining Graph Convolutional Networks (GCN) with resting-state functional magnetic resonance imaging (rs-fMRI) data to achieve early diagnosis of Autism Spectrum Disorder (ASD). The paper is co-authored by Chaoran Ma, Wenjie Li, Sheng Ke, Jidong Lv, Tiantong Zhou, and Ling Zou. It was published online on March 8, 2024, by the International Federation for Medical and Biological Engineering.

Research Background:

Autism Spectrum Disorder (ASD) is a heterogeneous condition characterized by repetitive behaviors, narrow interests, and severe social interaction deficits, meaning that it manifests differently in different individuals. The prevalence of autism among Chinese preschool children is about 1%. Currently, the diagnosis of autism relies on diagnostic scales and physician inquiries. This highly subjective assessment method significantly affects diagnostic outcomes and poses major challenges to healthcare, society, and educational care.

Functional magnetic resonance imaging (fMRI) is a high-resolution non-invasive imaging technique that has been widely used to detect and diagnose various brain diseases, including autism. In recent years, early diagnosis of ASD based on fMRI data using deep learning techniques has shown great potential. Blood oxygen level-dependent (BOLD) signals obtained from fMRI-based ASD studies can be used to construct functional connectivity (FC) and explore differences in brain activity and related biomarkers between patients and normal individuals. Brain network analysis using a graph structure has become one of the important methods for understanding brain function.

Paper Summary:

This paper proposes a multiple functional connectivity-based graph convolutional network (mfc-GCN) framework that uses not only whole-brain functional connectivity data but also functional connectivity data from key brain regions associated with ASD. It uses Graph Convolutional Networks (GCN) to capture complementary feature information for the final classification task. To address the heterogeneity issue in the ASD brain imaging data exchange (ABIDE) dataset, a novel external attention network readout layer (eanreadout) is designed and introduced to explore potential individual associations and effectively solve the heterogeneity problem in the dataset. The experimental results show that the framework can effectively learn complementary feature information from both types of data, achieving a final ASD classification accuracy of 70.31%.

Research Methods:

The study selected the ABIDE dataset of 714 subjects, covering 334 ASD patients and 380 healthy controls. Among them, there were 289 male patients, 45 female patients, 308 male healthy controls, and 72 female healthy controls. The brain was partitioned using Automated Anatomical Labeling (AAL) to obtain time series data, which was fed into the multiple functional connectivity-based graph convolutional network framework (mfc-GCN). The framework has two branches: one for feature learning from whole-brain FC data and the other for feature learning from key FC data. The features extracted from the two branches are concatenated for the final ASD classification task.

Graph Construction and Feature Extraction

Using AAL partitions, 116 brain regions were obtained, and Pearson correlation coefficients between time series were calculated to obtain functional connectivity data (whole-brain FC and key FC). To trim unnecessary information, the K-nearest neighbor algorithm (KNN) was introduced, retaining the top 10% strongest edges for each node when processing whole-brain FC and key FC.

Graph Convolutional Network (GCN)

GCN was used to perform feature learning on the input graph. GCN defines a convolution operation per layer to aggregate and update the nodes’ information from their neighboring nodes based on the convolution theorem of the Fourier transform of the graph data structure.

External Attention Network Readout Layer (EANReadout)

To address the data heterogeneity issue, a new external attention network readout layer (eanreadout) was designed, which learns potential correlations between samples to reduce heterogeneity problems. eanreadout effectively captures cross-subject information, improving the overall network performance and ASD diagnostic capability.

Main Results:

  1. Performance Comparison: The proposed mfc-GCN framework excels in accuracy, sensitivity, and specificity, with test results of 70.31%, 72.55%, and 69.23%, respectively.
  2. Comparison of Different Readout Layers: Comparing the performance of max, mean, and their concatenated readout layers with eanreadout, the eanreadout readout layer outperforms traditional readout layers in all metrics, with accuracy improved by 4.32%.
  3. Effectiveness of Multi-layer eanreadout Concatenation: The best results were achieved by concatenating three GCN layers (r1||r2||r3) during multiple eanreadout concatenation experiments.
  4. Effectiveness of Whole Brain and Key Functional Connectivity: mfc-GCN achieves the best performance by simultaneously using whole-brain and key FC data, demonstrating the complementarity of feature learning from both.

Conclusion and Significance:

The proposed multiple functional connectivity-based graph convolutional network (mfc-GCN) framework combines whole-brain FC and key FC data, using GCN to extract complementary feature information from both global and local topologies, enhancing the model’s ASD classification ability. The novel external attention network readout layer (eanreadout) captures cross-subject information, effectively addressing the data heterogeneity problem, and further improving classification performance and diagnostic efficacy. Overall, the paper significantly improves ASD diagnostic accuracy by mining complementary features from multiple functional connectivity data and utilizing an innovative readout layer, providing important insights for future early ASD identification.

Future Prospects:

Future research may consider incorporating multimodal information such as structural magnetic resonance imaging (sMRI). Combining deeper neuroscientific interpretability with deep learning results could further deepen the understanding of ASD mechanisms.